reproducibility

Why LLM-as-judge scores flip between identical runs

An LLM judge gate flipped between 0.79 and 0.82 on identical inputs. Sampling temperature, model drift and vague rubrics cause the jitter; k-sampling and a noise band fix it.